Update README.md
Browse files
README.md
CHANGED
|
@@ -53,8 +53,7 @@ outputs = model.generate(
|
|
| 53 |
|
| 54 |
DeepConf’s online stopping threshold is derived from a short warmup phase. You collect warmup trace confidences, then pass them into the generator to auto-derive the threshold for either DeepConf-low (aggressive) or DeepConf-high (permissive).
|
| 55 |
|
| 56 |
-
|
| 57 |
-
|
| 58 |
```python
|
| 59 |
from transformers import GenerationConfig
|
| 60 |
|
|
@@ -83,9 +82,7 @@ out = model.generate(
|
|
| 83 |
warmup_C = out.confidences.min(dim=1).values.tolist()
|
| 84 |
```
|
| 85 |
|
| 86 |
-
|
| 87 |
-
2) Online: pass warmup confidences to auto-derive threshold
|
| 88 |
-
|
| 89 |
```python
|
| 90 |
gen_cfg = GenerationConfig.from_model_config(model.config)
|
| 91 |
gen_cfg.enable_conf = True
|
|
|
|
| 53 |
|
| 54 |
DeepConf’s online stopping threshold is derived from a short warmup phase. You collect warmup trace confidences, then pass them into the generator to auto-derive the threshold for either DeepConf-low (aggressive) or DeepConf-high (permissive).
|
| 55 |
|
| 56 |
+
1. Warmup (num_return_sequences): collect per-trace confidences (C_t = min(step_confidences))
|
|
|
|
| 57 |
```python
|
| 58 |
from transformers import GenerationConfig
|
| 59 |
|
|
|
|
| 82 |
warmup_C = out.confidences.min(dim=1).values.tolist()
|
| 83 |
```
|
| 84 |
|
| 85 |
+
2. Online: pass warmup confidences to auto-derive threshold
|
|
|
|
|
|
|
| 86 |
```python
|
| 87 |
gen_cfg = GenerationConfig.from_model_config(model.config)
|
| 88 |
gen_cfg.enable_conf = True
|